Joerg Hiller
Feb 22, 2026 04:38
LangChain details how its Agent Builder memory system uses filesystem metaphors and COALA framework to create persistent, learning AI agents without code.
LangChain has pulled back the curtain on the memory architecture powering its LangSmith Agent Builder, revealing a filesystem-based approach that lets AI agents learn and adapt across sessions without requiring users to write code.
The company made an unconventional bet: prioritizing memory from day one rather than bolting it on later like most AI products. Their reasoning? Agent Builder creates task-specific agents, not general-purpose chatbots. When an agent handles the same workflow repeatedly, lessons from Tuesday’s session should automatically apply on Wednesday.
Files as Memory
Rather than building custom memory infrastructure, LangChain’s team leaned into something LLMs already understand well—filesystems. The system represents agent memory as a collection of files, though they’re actually stored in Postgres and exposed to agents as a virtual filesystem.
The architecture maps directly to the COALA research paper’s three memory categories. Procedural memory—the rules driving agent behavior—lives in AGENTS.md files and tools.json configurations. Semantic memory, covering facts and specialized knowledge, resides in skill files. The team deliberately skipped episodic memory (records of past behavior) for the initial release, betting it matters less for their use case.
Standard formats won out where possible: AGENTS.md for core instructions, agent skills for specialized tasks, and a Claude Code-inspired format for subagents. The one exception? A custom tools.json file instead of standard mcp.json, allowing users to expose only specific tools from MCP servers and avoid context overflow.
Memory That Builds Itself
The practical result: agents that improve through correction rather than configuration. LangChain walked through a meeting summarizer example where a user’s simple “use bullet points instead” feedback automatically updated the agent’s AGENTS.md file. By month three, the agent had accumulated formatting preferences, meeting-type handling rules, and participant-specific instructions—all without manual configuration.
Building this wasn’t trivial. The team dedicated one person full-time to memory-related prompting alone, solving issues like agents remembering when they shouldn’t or writing to wrong file types. A key lesson: agents excel at adding information but struggle to consolidate. One email assistant started listing every vendor to ignore rather than generalizing to “ignore all cold outreach.”
Human Approval Required
All memory edits require explicit human approval by default—a security measure against prompt injection attacks. Users can disable this “yolo mode” if they’re less concerned about adversarial inputs.
The filesystem approach enables portability that locked-in DSLs can’t match. Agents built in Agent Builder can theoretically run on Deep Agents CLI, Claude Code, or OpenCode with minimal friction.
What’s Coming
LangChain outlined several planned improvements: episodic memory through exposing conversation history as files, background memory processes running daily to catch missed learnings, an explicit /remember command, semantic search beyond basic grep, and user-level or org-level memory hierarchies.
For developers building AI agents, the technical choices here matter. The filesystem metaphor sidesteps the complexity of custom memory APIs while remaining LLM-native. Whether this approach scales as agents handle more complex, longer-running tasks remains an open question—but LangChain’s betting that files beat frameworks for no-code agent building.
Image source: Shutterstock
